Wireless relays are such methods that can be rapidly deployed to enhance the coverage, reliability and throughput of a wireless network subject to power and spectral constraints. When equipped with multiple antennas, Multiple Input Multiple Output (MIMO) relays, are particularly useful for scattering rich and non-line-of-sight environment. The first part of this dissertation considers a system where two users exchange information via a non-regenerative half-duplex two-way MIMO relay. We study the transceiver design including both source covariance matrices at the two users and beamforming matrix at the relay to maximize the achievable weighted sum rate of the system. We compare the convergence behaviors of the proposed algorithms and demonstrate their advantages over prior algorithms. We also show an optimal structure of the relay matrix, which is useful to reduce the search complexity.
As advanced architectures and high-performance hardware are required to implement more powerful but complicated algorithms such as those in a two-way relay system, multiple cores are often integrated on a chip of shrinking size. However, the corresponding dramatically increased power density may lead to significant adverse effects. Dynamic thermal management is widely used to mitigate this problem, where thermal modeling and temperature prediction play the key roles. Unlike conventional bottom-up approaches, a Linear Time-Invariant (LTI) Multiple-Input-Multiple-Output (MIMO) black-box model is adopted and Least Square (LS) based model averaging algorithm with model screening is developed with less temperature prediction errors than the traditional LS algorithm based on model order selection.
For VLSI thermal modeling, Model Order Reduction (MOR) is an efficient technique to reduce the modeling complexity where subspace-based methods could be successfully applied. In the third part, motivated by MOR for thermal modeling, subspace tracking is investigated as one of the key procedures for subspace-based methods. We explain the reason to enlarge the actual tracking dimension and equip bi-iteration SVD algorithm with multiple inner iterations and Ritz acceleration. Our proposed algorithms are demonstrated to have much improved performance of both convergence rate and tracking accuracy compared to existing algorithms while still keeping linear complexity without many additional computational consumptions.
